self-healing-claudex

An autonomous build-and-review loop pairs Claude-based planning and building with Codex-based adversarial review to converge on a sound build. It coordinates planning, building, and multi-round reviews until no substantive findings remain or maximum rounds are reached.

Adversarial code review uses Codex-driven deterministic personas to surface actionable findings on code changes and test contracts. It enforces an explicit spec-claims contract and PLAN.md tests to ensure coverage gaps are checked before a build is accepted.

You automate multi-round code reviews by orchestrating a self-healing loop that applies atomic state writes and round transitions. It tracks stateful progress, enforces safety guards, and surfaces summarized outcomes for user review after each cycle.

Yes, the loop enforces spec-claims contracts and validates existing test contracts defined in PLAN.md. It coordinates builder execution with Codex adversarial review to identify coverage gaps and ensure test integrity across code changes.

The best way is to run an autonomous dual-review loop that cycles through a Claude-driven planner and a Codex-driven adversarial reviewer. This iteratively validates code changes until substantive findings are exhausted or max rounds are reached.

The loop stops when maximum rounds are reached, even if substantive findings remain. It enforces this constraint through lifecycle state management with atomic state writes and round transitions, then surfaces a final summary of unresolved findings.